TechLook
The infrastructure bet is so strong that even America's spies can't get in the door. That's the through-line of today's news — and it's more revealing than any product launch.
America's Spies Are Paying $9B to Wait Behind a Chatbot
The White House quietly approved a $9 billion secret budget — plus $800M reprogrammed — so the CIA and NSA can finally buy Nvidia Grace Blackwell chips to run frontier AI on classified networks. The problem: the latest ChatGPT and Anthropic's Mythos model won't fit on air-gapped machines. The agencies that used to invent the future can't log into it.
The Pentagon had labeled Anthropic a supply chain threat. The White House overrode them and waived the "any lawful use" clause every other vendor signed — just to keep NSA access alive.
The inversion is total. For seventy years, spies got the toys first, the rest got leftovers a decade later. Mythos shipped to JPMorgan and British banks months before Langley could host it. Even with $9B wired, the NSA waits behind a hedge fund's risk desk.
Here's everything you need to know:
- $9B in secret funding approved for CIA/NSA to buy Nvidia Grace Blackwell chips
- $800M additional reprogrammed for AI infrastructure on classified networks
- Pentagon had flagged Anthropic as a supply chain threat — White House overrode that designation
- Anthropic waived the "any lawful use" clause that every other vendor signed
- Mythos shipped to JPMorgan and British banks months before US intelligence agencies could access it
- CIA and NSA can't run frontier models on air-gapped systems — the hardware and architecture don't fit
The vendor the Pentagon flagged as a threat just got promoted to writing the procurement rules. America's intelligence community is writing a $9B check and still standing in line.
AI Memory Hit a Wall — And the Market Noticed
Micron briefly crossed $1 trillion market cap Tuesday on a 19% pop, joining Samsung in the club. But the number is just the surface. What's underneath is more instructive: Micron's entire 2026 HBM supply is sold out. HBM4 is already in production. Customers are signing multi-year contracts at fixed prices.
Every generation of frontier model has eaten more memory than the last. Agents bent the curve vertical. Long-context inference, KV caches, and persistent agent state burn bandwidth and capacity at multiples of pure inference. When an entire calendar year of HBM is sold out before the chips ship — and Anthropic is rationing tokens to paying users — demand has stopped looking like a market and started looking like infrastructure draw.
Here's everything you need to know:
- Micron briefly crossed $1T market cap on 19% pop — Samsung already in the club
- SK Hynix also crossed $1T for the first time this week
- Roundhill Memory ETF (DRAM) became the fastest ETF in history to reach $10B in assets
- Micron's entire 2026 HBM supply is sold out
- HBM4 already in production with multi-year customer commitments
- Anthropic is rationing tokens to paying users — demand outstrips supply at the model layer too
The memory layer is the new bottleneck. If you're building anything that involves long context, agentic workflows, or persistent state, the constraint isn't your code — it's the hardware that isn't being made fast enough.
Stanford Found Systemic Bias in AI Hiring Tools — And the Fix Isn't Simple
Stanford researchers analyzed 4 million job applications across 156 employers using Pymetrics AI hiring tools and found what they called "clear racial disparities." 10.62% of positions showed adverse impact against Black applicants. 5.32% against Asian applicants. The problem isn't just the bias — it's the architecture: 42 shared models across employers, meaning rejection at one company can cascade to others using the same model. 4% of applicants who applied to 10 positions were rejected from all of them — higher than random chance would predict.
The data covers 2018–2022, so modern LLM-based hiring tools may work differently. But the shared-infrastructure risk doesn't care about the vintage of the model.
Here's everything you need to know:
- 10.62% of positions showed adverse impact against Black applicants
- 5.32% of positions showed adverse impact against Asian applicants
- 42 shared models across employers — rejection at one company can cascade to others
- 4% of applicants who applied to 10 positions were rejected from all of them
- Data covers 2018–2022 (Pymetrics-era tools), may not generalize to current LLM-based systems
- Shared infrastructure means the bias is systemic, not company-specific
The talent industry is racing to layer LLMs on top of these systems. If you're hiring and using any vendor that shares model infrastructure across clients, you're inheriting someone else's bias problem. Know what model is running your pipeline.
The Dropbox Founder Left to Go Build AI — And That Says Everything
Drew Houston is stepping down as CEO of Dropbox after 19 years. He told CNBC there's no "SaaS apocalypse" and no customer is canceling Dropbox because of ChatGPT. In the same breath, he announced his next chapter: building something in AI — just not at Dropbox.
The cloud storage pioneer is worth $6B today, half its 2018 IPO peak and below its 2014 private valuation of $10B. Revenue has been flat for two years. Asana, Monday.com, and HubSpot are all down more than 60% in a year. The SaaS incumbents have spent 18 months telling the market AI is a tailwind, not a wrecking ball. Their stocks disagree.
Here's everything you need to know:
- Drew Houston stepping down after 19 years as Dropbox CEO
- Stock at $6B — half its 2018 IPO peak, below its 2014 private valuation of $10B
- Revenue flat for two years
- Asana, Monday.com, HubSpot all down more than 60% in a year
- Houston: "no SaaS apocalypse," customers not canceling for ChatGPT — then announced he's leaving to build AI elsewhere
Houston spent the morning telling reporters AI isn't killing SaaS. Then he announced he was leaving to go build AI somewhere SaaS isn't dying. The founder's feet said what his press tour wouldn't.
When PR Firms Start Refusing Pitches on the Record, the Label Has Outrun the Substance
UK PR firms say their clients are forcing them to pitch every product as AI. AllBirds "pivoted" to acquiring AI GPUs. Inboxes are carrying releases for AI basketball hoops and AI lasers that protect women on subway platforms. One agency director said half the AI stories he sends out, he wishes he didn't.
For a decade the label of choice was "data-driven" or "platform." AI ate both. Any company without it is invisible to the board and the buyside — so a shoe brand stockpiles GPUs and a handheld floor scanner gets relabeled AI-powered.
Here's everything you need to know:
- UK PR firms refusing to pitch AI products out of "professional shame"
- AllBirds "pivoted" to acquiring AI GPUs
- AI basketball hoops, AI subway safety lasers now being pitched
- One agency director: half the AI stories he sends out, he wishes he didn't
- AI replaced "data-driven" and "platform" as the required label
- PR industry sold NFTs without flinching — now refusing AI pitches on the record
When the industry that sold NFTs without blinking starts refusing AI pitches on the record, that's a sell signal. Not on AI — on the label.
Exa Labs Raised $250M to Own the Infrastructure Underneath AI Search
The future of online search is splitting into two bets. On one side: Exa Labs raised $250M at a $2.2B valuation building search infrastructure specifically for AI agents — not humans. Parallel Web Systems has raised hundreds of millions. Tavily and TinyFish are close behind. On the other side: Google announced AI-powered Search agents rolling out this summer — its biggest redesign in 25 years.
The bind is real: agents don't click ads. An agent-first internet could put Google's $200B+ Search business in jeopardy. Whoever wins the transformation will likely own the next era of the internet.
Here's everything you need to know:
- Exa Labs raised $250M at $2.2B valuation building agent-first search infrastructure
- Parallel Web Systems, Tavily, TinyFish all raising/investing heavily
- Google redesigning Search with AI agents rolling out this summer
- Google's biggest redesign in 25 years
- Agents don't click ads — conflict between agent-first web and Google's ad revenue model
- Hundreds of millions flowing to startups racing to own infrastructure underneath
If you're building anything that involves information retrieval, you need to decide which world you're building for: humans who click, or agents who read.
Jensen Huang Says Stop Chasing "AI-Proof" Careers
Nvidia CEO Jensen Huang advised parents and students not to chase "AI-proof" career paths. Speaking with CNA, he said students should ask: "How can AI help elevate my learning, my craft, my purpose?" He used journalism as an example — the best don't just prepare questions, they listen, think about the audience, and respond dynamically. He referenced "wabi-sabi" — the beauty of imperfection — suggesting uniquely human qualities become more prized.
He called the narrative linking AI to job cuts "lazy": "AI has just arrived, how is it possible they're already losing jobs?"
Here's everything you need to know:
- Jensen Huang: don't chase "AI-proof" careers — ask how AI elevates your craft
- Used journalism as example: best listen, think about audience, respond dynamically
- Referenced "wabi-sabi" — beauty of imperfection
- Called AI-to-job-cuts narrative "lazy"
- Context: 80K+ jobs already cut this year despite Huang's optimism
The philosophy is clean. The hiring data is messier. But the direction is right: taste, original thinking, and emotional connection become more valuable, not less, as AI handles the rest.
⚡ Quick Hits
-
Anthropic's Mythos: Solved Erdős Problem #90 — same result as OpenAI but with a simpler proof. Research progress, not just benchmark warfare.
-
Xiaomi: Cut MiMo-V2.5 series API pricing by up to 99% and increased token allowances 5–8x. The pricing war in small models continues to compress margins.
-
OpenRouter: Raised $113M, scaled to 8M developers with 1.5 quadrillion-token annual run rate. The model aggregation layer is becoming infrastructure.
-
ElevenLabs: Released Music v2 with better vocals, instrumentation, multilingual support, and track-level inpainting. Generative audio keeps advancing.
-
China: Imposing overseas travel restrictions on top AI researchers from Alibaba and DeepSeek. The talent mobility crackdown is real and accelerating.
-
Jensen Huang: Said taste, original thinking, and emotional connection become more valuable, not less, as AI handles the rest. Worth writing down.
-
Dropbox: Founder exits after 19 years to "build something in AI" — stock down 50% from 2018 IPO peak. The SaaS incumbents have a credibility problem.
Techlook — AI & tech signal for founders and builders.